SIMULATED — WORKSHOP DEMO · synthetic borrower · no live RBIH/ULI/AA call · every fact is sourced and labelled
100% · 21/22
Close / governance

Survive the inspection because the behavior is legible.

The demo is not asking the room to trust a model. It maps each behavior to the operative draft MRM expectations and keeps the accountable human in the loop.

Instrument status:
Point-by-point mapping

Draft MRM guidance → demo behavior

Mapped means the product behavior is visible in this workshop. It does not mean a bank’s board framework or validation obligation has been discharged.

1
Board-approved MRM framework
ACCOUNTABLE · Bank MRM owner

The demo exposes a named control register, human authority boundary and exception path; the bank owns the framework and approval.

Mapped
2
Independent validation
ACCOUNTABLE · Independent validation

Every model entry has a validation owner and date. Validation is separate from the agent builder, with adverse findings retained.

Mapped
3
Mandatory human oversight for material decisions
ACCOUNTABLE · Credit officer / committee

The facility slider is a logged human action. Conflict, sanction, fraud classification and limit movement cannot be completed by the model.

Live boundary
4
Explainability for customer-facing models
ACCOUNTABLE · Product + credit risk

Claims carry source, artefact, timestamp, transformation and confidence. The borrower view removes internal scores and jargon.

Mapped
5
Disclosure that the user is interacting with AI
ACCOUNTABLE · Product owner

The persistent workshop banner and borrower-facing copy disclose scripted AI behavior before a data pull or recommendation.

Visible
6
Third-party and vendor accountability
ACCOUNTABLE · RE / TSP governance

The provenance surface records model, agent and vendor version. The regulated entity remains accountable for its LSP.

Mapped
7
Adversarial and edge-case testing
ACCOUNTABLE · Model risk + credit

The adversarial agent tries to prevent sanction; source conflict and stale land records are carried forward rather than smoothed.

Demonstrated
8
Model inventory and versioning
ACCOUNTABLE · Model risk office

The live inventory below records purpose, owner, version, validation date and known failure mode, including the land-use retraction.

Visible
Model risk register

Model and agent inventory

SIMULATED WORKSHOP REGISTER
Model / agentVersionOwnerLast validationKnown failure mode
fabric-research0.9.3Research lead11 Aug 2026Missed or stale source; conflict is carried forward
evidence-reconcile0.8.1Credit risk11 Aug 2026Cannot resolve 18.8% turnover variance without officer
land-retrieval0.6.4Operations risk11 Aug 2026Stale district extract; retracted after later revenue record
ews-reconciler0.7.0Portfolio risk11 Aug 2026False positive from isolated signal; assumption link required
adversarial-credit0.4.2Independent validation11 Aug 2026May over-index on downside; cannot sanction
Retraction is an inventory event: the land-use retrieval model’s stale assertion is not deleted. Its original claim, source, challenger, corrected source and downstream CP change remain in the audit trail.
Human authority

AI recommends. The human decides.

A model can surface a risk, draft a CAM or propose a pathway. It cannot move a credit limit, resolve a material conflict or classify fraud.


TSP-neutral position

Make the bank’s existing core intelligent.

The next generation of TSP is not the company that replaces the bank’s core. It is the company that makes the bank’s existing core intelligent. This layer stays TSP-neutral.


That accountability is why provenance has to be a product surface, not a promise.
The bank does not outsource accountability by adding intelligence. It makes accountability inspectable at every hand-off.